A unit business record set generation method and device based on a large model

Abstract

The application discloses a kind of unit business chronicle set generation method and device based on large model, which comprises the following steps: S1, obtains unit organization structure and function division information and unit business material set;S2, unit organization structure and function division information are processed, and unit business classification system and unit business classification label library are obtained;S3, based on unit business classification system and unit business classification label library, unit business material set is processed, and training data set and verification data set are obtained;S4, using training data set and verification data set, the preset data analysis model is trained and verified, and unit business chronicle set generation model is obtained;S5, using unit business chronicle set generation model, the unit business data of target year is processed, and the unit business chronicle set of target year is obtained.The method of the application is suitable for government agencies, institutions and enterprises, and can automatically and accurately generate chronicle set reflecting important activities and development history of the unit.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and artificial intelligence, and in particular to a method and apparatus for generating unit business record sets based on a large model. Background Technology

[0002] An organization's operational chronicle is a standardized document that records an organization's important work, major events, and key achievements within a specific period. It is of great significance for historical tracing, work review, achievement presentation, and external communication. Traditionally, the generation of organizational chronicles relies heavily on manual compilation. Staff must sift through a vast amount of offline documents, online news, and business records based on pre-defined keywords, manually organizing the timeline, extracting core content, and arranging it according to specifications. However, traditional methods have several drawbacks: First, they are inefficient. Screening and extracting information from massive amounts of documents requires a significant amount of manpower and time, especially for organizations with complex business operations and abundant materials. Manual compilation is time-consuming and susceptible to fatigue. Second, they lack accuracy. Manual screening can easily lead to omissions of key information, errors in timelines, and deviations in the extraction of core content. Furthermore, inconsistent standards among different compilers make it difficult to ensure the standardization and uniformity of the organization's business chronicles. Third, they have limited coverage. Manual methods cannot comprehensively cover materials from all business areas, which may result in some important business-related events not being included in the organization's business chronicles. Fourth, they lack dynamic adaptability. As an organization's business areas may change with development, traditional methods cannot quickly adapt to these changes.

[0003] Therefore, there is an urgent need for a method to generate a set of business records for a unit, in order to overcome the above-mentioned shortcomings. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and apparatus for generating unit business record sets based on a large model, thereby solving the technical problems of low efficiency, insufficient accuracy, limited coverage and lack of dynamic adaptation capability in the existing technology of manually compiling and producing unit business record sets.

[0005] To address the aforementioned technical problems, the first aspect of this invention discloses a method for generating unit business record sets based on a large model, the method comprising: S1, obtain information on the organization's organizational structure and functional division of labor, as well as the organization's business materials; S2, Process the organizational structure and functional division information of the unit to obtain the unit business classification system and the unit business classification tag library; S3. Based on the unit business classification system and the unit business classification label library, the unit business material set is processed to obtain a training dataset and a validation dataset. S4. Using the training dataset and the validation dataset, the preset data analysis model is trained and validated to obtain the unit business record generation model. The data analysis model includes four independent task-customized neural network models: a time information extraction sub-model, an entity information extraction sub-model, an event information extraction sub-model, and an image association and information source extraction sub-model. Each sub-model includes at least an independent feature extraction layer and an output layer. S5. Using the aforementioned unit business record set generation model, process the unit business data for the target year to obtain the unit business record set for the target year.

[0006] As an optional implementation, in the first aspect of the present invention, processing the organizational structure and functional division information of the unit to obtain a unit business classification system and a unit business classification tag library includes: S21, The organizational structure and functional division information of the unit are parsed and processed to obtain the unit's business information; S22, The unit business information is processed using a hierarchical classification method to obtain an initial unit business classification system; S23, The initial unit business classification system is modified to obtain the unit business classification system; S24. Based on the unit business classification system, process the unit business information to obtain a unit business classification tag library.

[0007] As an optional implementation, in the first aspect of the present invention, the step of processing the unit business material set based on the unit business classification system and the unit business classification label library to obtain a training dataset and a validation dataset includes: S31, Establish a unit business material index library based on the aforementioned unit business material set; S32, Based on the unit business material index library, perform data extraction processing on the unit business material set to obtain the unit business dataset; S33, Based on the unit business classification system and the unit business classification label library, the unit business dataset is annotated with relevant information of the record set to obtain the unit business labeled dataset; S34, Select and divide the unit business annotation dataset to obtain the training dataset and the validation dataset.

[0008] As an optional implementation, in the first aspect of the present invention, the step of training and validating a preset data analysis model using the training dataset and the validation dataset to obtain a unit business record generation model includes: S41, using the training dataset, train the time information extraction sub-model, the entity information extraction sub-model, the event information extraction sub-model, and the image association and information source extraction sub-model respectively, to obtain the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model; S42, the Transformer is used to fuse and collaboratively optimize the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model to obtain a multi-task collaborative data analysis model; S43, using the unit business classification tag library, train the business domain matching capability of the multi-task collaborative data analysis model to obtain the first unit business record generation model; S44, perform text data augmentation training on the first unit business record generation model to obtain the second unit business record generation model; S45, perform adversarial optimization training on the second unit business record generation model to obtain the third unit business record generation model; S46, perform cross-format text adaptation training on the third unit business record generation model to obtain the fourth unit business record generation model. S47, Using the verification dataset and preset multi-dimensional verification indicators, a verification experiment is conducted on the fourth unit business record generation model to obtain the compliance verification result; When the compliance verification result is yes, the fourth unit business record set generation model is determined to be a unit business record set generation model; If the compliance verification result is negative, repeat steps S43 to S47 until the compliance verification result is positive.

[0009] As an optional implementation, in the first aspect of the present invention, training the event information extraction sub-model using the training dataset to obtain a corresponding optimized event information extraction sub-model includes: S411, Construct an event type and trigger word mapping library based on the unit's business material set; S412, using the event type and trigger word mapping library, perform event trigger word annotation, event type annotation and event status annotation processing on the training dataset to obtain an event annotation sample set; S413, Perform trigger word context enhancement and sample balancing on the event-labeled sample set to obtain the event training sample set; S414, using the event type training weight calculation model and the event training sample set, the event information extraction sub-model is jointly fitted and trained with event type and event state to obtain the event information extraction sub-model to be optimized; the expression of the event type training weight calculation model is: In the formula, The normalized training weights for the k-th class of event samples take values ​​in the range (0,1). Let be the normalized frequency of the k-th type of event, with a value range of (0,1]. This is a preset frequency enhancement factor, with a value range of [0.3, 2.0]. Let be the sample difficulty coefficient for the k-th type of event. The initial prediction confidence of the event information extraction sub-model for the k-th type of event sample is given, with a value range of [0,1). Let be the rarity coefficient of the k-th type of event. Let k be the number of samples of the k-th type of event. The total number of samples in the training sample set for the event. The preset rarity compensation factor has a value range of [0.3, 2.0]; K is the total number of event types in the event training sample set; S415, Perform time and event state correlation verification and feature constraint optimization on the event information extraction sub-model to be optimized to obtain the event information extraction optimization sub-model.

[0010] As an optional implementation, in the first aspect of the present invention, the step of using Transformer to fuse and collaboratively optimize the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model to obtain a multi-task collaborative data analysis model includes: S421, the feature extraction layers of the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model are integrated using Transformer to obtain a feature sharing layer; S422, taking the feature sharing layer as the core, the output layer of the time information extraction optimization sub-model, the output layer of the entity information extraction optimization sub-model, the output layer of the event information extraction optimization sub-model, and the output layer of the image association and information source extraction optimization sub-model are associated to obtain the first multi-task collaborative data analysis model; S423, with the goal of minimizing the cross-task loss function, the first multi-task collaborative data analysis model is trained with cross-task collaborative constraints using the training dataset to obtain the second multi-task collaborative data analysis model; The expression for the cross-task loss function is: In the formula, This represents the total loss across tasks, and its value ranges from [0, +]. ); , , , These are the dynamic weight coefficients for the time information extraction subtask, the entity information extraction subtask, the event information extraction subtask, and the image association and information source extraction subtask, respectively. The values ​​are all in the range of (0,1] and the sum of the weights is 1. , , , These are the training errors for the time information extraction subtask, entity information extraction subtask, event information extraction subtask, and image association and information source extraction subtask, respectively, all ranging from [0, +]. ); The preset logical constraint penalty coefficient has a value range of [0, 3, 0.7]. For indicator functions; This is a multi-task logical consistency judgment function; The structured prediction output of the sub-model is optimized for extracting time information. The structured prediction output of the sub-model is optimized for entity information extraction; Optimize the structured prediction output of the sub-model for event information extraction; The structured prediction output of the optimized sub-model for image association and information source extraction.

[0011] S424, perform iterative collaborative optimization processing on the second multi-task collaborative data analysis model to obtain the multi-task collaborative data analysis model.

[0012] As an optional implementation, in the first aspect of the present invention, the step of using the unit business classification tag library to train the multi-task collaborative data analysis model to perform business domain matching capability training to obtain a first unit business chronicle generation model includes: S431, Perform business domain label normalization processing on the unit business classification label library to obtain a standardized business label set; S432, Perform business domain association matching processing on the standardized business label set and the training dataset to obtain a business domain matching training sample set; S433, using the business domain matching training sample set, the multi-task collaborative data analysis model is trained to perform business semantic adaptation and domain feature fitting to obtain the multi-task analysis model to be adapted. S434, Perform business classification label association constraint and domain matching accuracy calibration on the multi-task analysis model to be adapted to obtain the first unit business record set generation model.

[0013] A second aspect of this invention discloses a unit business record set generation device based on a large model, the device comprising: The unit information acquisition module is used to acquire information on the unit's organizational structure and functional division of responsibilities, as well as the unit's business materials. The unit business classification module is used to process the unit's organizational structure and functional division information to obtain the unit business classification system and the unit business classification tag library; The unit business material processing module is used to process the unit business material set based on the unit business classification system and the unit business classification label library to obtain a training dataset and a validation dataset. The unit business history generation model training module is used to train and validate a preset data analysis model using the training dataset and the validation dataset to obtain the unit business history generation model. The unit business history generation module is used to process the unit business data of the target year using the unit business history generation model to obtain the unit business history of the target year.

[0014] A third aspect of this invention discloses a unit business record set generation device based on a large model, the device comprising: processor; A memory coupled to the processor stores executable program code; The processor calls the executable program code stored in the memory to execute the unit business record generation method based on a large model disclosed in the first aspect of the present invention.

[0015] A fourth aspect of the present invention discloses a computer-readable storage medium storing computer instructions, which, when invoked, are used to execute the unit business record set generation method based on a large model disclosed in the first aspect of the present invention.

[0016] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention fully utilizes the information extraction, analysis, and generation capabilities of large models in natural language processing. Through comprehensive material collection, standardized business classification, standardized generation of training and validation data, and multi-dimensional training, a unit business chronicle generation model is obtained. This unit business chronicle generation model can solve the technical problems of low efficiency, insufficient accuracy, limited coverage, and lack of dynamic adaptation capabilities in traditional manual generation of unit business chronicles, thereby achieving efficient, accurate, and standardized generation of unit business chronicles. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating a method for generating a unit business record set based on a large model, as disclosed in an embodiment of the present invention. Figure 2 This is a schematic diagram of a unit business record set generation device based on a large model disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of another unit business record set generation device based on a large model disclosed in an embodiment of the present invention; Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The terms "first," "second," etc., used in the specification and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] In this application, the term "exemplary" is used to mean "used as an example, illustration, or description." Any embodiment described as "exemplary" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes are not described in detail to avoid obscuring the description of this application with unnecessary detail. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0023] It should be noted that since the method in this application embodiment is executed in a computer device, the processing objects of each computer device exist in the form of data or information, such as time, which is essentially time information. It is understood that if size, quantity, position, etc. are mentioned in subsequent embodiments, they are all corresponding data that exist so that the computer device can process them. Specific details will not be elaborated here.

[0024] It should be noted that the artificial intelligence-related technologies that may be involved in this application will be briefly described. Artificial intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. Artificial intelligence is the study of the design principles and implementation methods of various intelligent machines, enabling machines to have the functions of perception, reasoning, and inference output.

[0025] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0026] In all embodiments of the present invention, the variables involved in all computational expressions or mathematical functions are dimensionless before calculation. The values ​​of the independent variables in all computational expressions or mathematical functions in these embodiments conform to the reasonable requirements of the input range of the computational expression or mathematical function, ensuring that the computational expression or mathematical function can be calculated smoothly without violating physical laws or mathematical rules.

[0027] This application provides a method and apparatus for generating unit business record sets based on a large model, which will be described in detail below.

[0028] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for generating unit business record sets based on a large model, as disclosed in an embodiment of the present invention. Figure 1 The described method for generating unit business timelines based on a large model is applied in a unit business timeline generation device, such as a local server or cloud server used for unit business timeline generation, etc., and the embodiments of the present invention are not limited thereto. Figure 1 As shown, a method for generating unit business record sets based on a large model includes: S1, obtain information on the organization's organizational structure and functional division of labor, as well as the organization's business materials; It should be noted that the above-mentioned organizational structure and functional division information refers to structured information reflecting the organization's overall organizational system, departmental setup, job allocation, affiliation, boundaries of authority and responsibility, and scope of business functions, specifically including: Organizational structure information: organizational hierarchy, internal departments, branches, staffing levels, personnel affiliations, reporting relationships, and departmental functions; Functional division of labor information: the scope of responsibilities, business authority, work content, areas of responsibility, boundaries of responsibility, business affiliation, and work process nodes of each department and position; It should be noted that the aforementioned collection of organizational business materials refers to the collection of various original materials generated, retained, and archived by the organization in its daily performance of duties, business operations, and management activities. This includes text, charts, images, meeting minutes, official documents, reports, briefings, summaries, ledgers, notices, approvals, activity materials, and videos, recording all important business matters, work results, major activities, important meetings, key events, important decisions, and performance of duties that occurred within the organization's annual, quarterly, and monthly periods. It includes both online and offline business materials. Specifically: Offline business materials are original materials formed in paper, physical media, or on-site recording methods. They are direct records of the organization's historical business activities and possess the characteristics of authenticity, originality, and traceability. They are the core authoritative data source for constructing the organization's business material collection and supporting the generation of the organization's business chronicle collection. They mainly include paper documents, official documents, notices, approvals, letters, meeting minutes, work briefings, work summaries, work reports, meeting minutes, sign-in sheets, on-site records, project materials, contracts, ledgers, reports, vouchers, archives, activity photos, promotional materials, approval documents, handwritten records, etc. The acquisition method is as follows: First, from physical storage locations such as archives, data rooms, warehouses, departmental file cabinets, and on-site work record books, they are searched, consulted, screened, and organized by year, department, and business category. Then, the screened paper materials are photocopied, photographed, scanned, OCR text recognition, electronically entered, formatted, deduplicated, and verified. Finally, after compliance review, confidentiality review, and content verification, they are uniformly archived and stored in the database to form structured electronic archives and included in the organization's business material collection. Online business materials are electronic documents generated, circulated, stored, and archived through digital channels such as information-based office work, digital management, online business processing, e-government, video conferencing, mobile office, cloud storage, business systems, and management platforms. They also include press releases, notices, announcements, and work updates published on the organization's official website; news related to the organization reported by external media; and relevant information published on industry platforms. These electronic materials and related information record daily business, important meetings, major events, key decisions, work updates, and cross-departmental collaborations in real time. They are characterized by high timeliness, comprehensive coverage, structured data, and automatic collection capabilities, making them an important data source for building high-quality, timely business material sets. The acquisition of online electronic materials mainly relies on the organization's information infrastructure, data interfaces, access control systems, and security. The control mechanism employs technologies such as automated data collection, batch synchronization, structured extraction, compliant web scraping, interface integration, and data cleaning. Specifically: First, through authorization, account authentication, secure interface calls, system integration, and data synchronization, electronic materials are automatically or semi-automatically exported, collected, and crawled from OA systems, document systems, meeting systems, business platforms, cloud storage, email, work groups, mobile office terminals, and internet platforms in batches. Then, the collected electronic materials undergo format standardization, deduplication, noise reduction, structured parsing, content verification, confidentiality filtering, and abnormal data removal. Finally, after security and compliance review, data anonymization, and access control, the materials are uniformly stored, categorized, archived, and structured, incorporated into the unit's business materials collection, and merged with offline electronic materials to form a complete, authoritative, and traceable data source.

[0029] S2, Process the organizational structure and functional division information of the unit to obtain the unit business classification system and the unit business classification tag library; S3. Based on the unit business classification system and the unit business classification label library, the unit business material set is processed to obtain a training dataset and a validation dataset. S4. Using the training dataset and the validation dataset, the preset data analysis model is trained and validated to obtain the unit business record generation model. The data analysis model includes four independent task-customized neural network models: a time information extraction sub-model, an entity information extraction sub-model, an event information extraction sub-model, and an image association and information source extraction sub-model. Each sub-model includes at least an independent feature extraction layer and an output layer. It should be noted that the above training and verification aims to conduct a full-process training and optimization of the preset data analysis model through supervised learning, from specialized training, model fusion, capability enhancement to effect verification, and finally obtain a dedicated model that can automatically, accurately and stably generate the unit's business record set.

[0030] It should be noted that the aforementioned pre-defined data analysis model consists of four independent functional sub-models, each of which includes at least an independent feature extraction layer and an output layer. Specifically: The time information extraction sub-model is used to accurately identify, extract, and structure and output time information such as year, month, date, time period, document writing time, activity time, and shooting time from unit business materials. It can adopt structures such as convolutional neural network, recurrent neural network, lightweight Transformer, or basic fully connected neural network. The embodiments of this invention are not limited to these structures. The entity information extraction sub-model is used to identify, locate and extract key entity information such as departments, personnel, locations, projects, organizations, document numbers, and cooperating units from the unit's business materials, clarifying the subject, object and responsibility elements of the unit's business chronicle. It can adopt structures such as convolutional neural networks, bidirectional recurrent neural networks, lightweight Transformers or conditional random fields, and the embodiments of this invention are not limited thereto. The event information extraction sub-model is used to identify, classify and extract event types such as meetings, activities, decisions, projects, commendations, documents, and approvals from the unit's business materials, as well as the status of events such as initiation, progress, and completion, to provide core event content for the unit's business chronicle. It can adopt structures such as convolutional neural networks, recurrent neural networks, lightweight Transformers or graph neural networks, and the embodiments of this invention are not limited thereto. The image association and information source extraction sub-model is used to establish the association between business text and image materials, and extract traceability information such as the source, collection department, storage path, and credibility rating of image metadata, so as to ensure that the text and images in the unit's business chronicle are consistent, the source is traceable, and the content is verifiable. It can adopt structures such as convolutional neural network, lightweight Transformer, multilayer perceptron or cross-modal fusion network, and the embodiments of this invention are not limited.

[0031] S5. Using the aforementioned unit business record set generation model, process the unit business data for the target year to obtain the unit business record set for the target year.

[0032] It should be noted that the target year can be a specific year or a specific time span, and the embodiments of the present invention do not limit it; the unit business data of the target year is input into the trained unit business chronicle generation model to automatically generate a target year unit business chronicle that is structurally standardized, content-complete, element-rich, consistent in definition, and fits the business scenario.

[0033] It should be noted that, in order to further improve the compilation quality of the aforementioned target year's unit business chronicle and the generation accuracy of the aforementioned unit business chronicle generation model, the aforementioned target year's unit business chronicle can be reviewed, focusing on checking the accuracy of the entry information, logical coherence, format standardization, and content completeness, and issuing rectification opinions; the unit business chronicle generation model adaptively adjusts its processing logic based on the review opinions, iteratively optimizes and outputs a revised version of the unit business chronicle; the closed-loop process of "review - model adjustment - iterative generation" is executed cyclically until all compilation requirements of the unit business chronicle are met.

[0034] It should be noted that the above-mentioned unit business record generation model can also be connected to the unit's information office network. According to the business processes and document management specifications of the office network, it can generate multi-format versions of the target year's unit business record, assign a unique identifier, and store it in the unit's business work database to achieve automated archiving and accurate retrieval and traceability, ensuring that the unit's business record is traceable and controllable throughout its entire lifecycle.

[0035] As can be seen, the unit business record set generation method based on a large model described in the embodiments of the present invention constructs a complete technical process from basic data collection, business system and business tag library construction, structured processing of training and validation data, multi-task model training to automatic generation of unit business record sets. This realizes the intelligent and automated generation of unit business record sets, effectively solving the problems of low efficiency, easy omission of information, inconsistent standards, and difficulty in tracing the source of traditional manual unit business record set compilation. It significantly improves the efficiency and quality of unit business record set compilation, ensures that the content is true and accurate, logically clear, and traceable, and provides efficient and reliable technical support for unit archives, work summaries, and work reviews.

[0036] In another optional embodiment, the processing of the organizational structure and functional division information of the unit to obtain the unit's business classification system and the unit's business classification tag library includes: S21, The organizational structure and functional division information of the unit are parsed and processed to obtain the unit's business information; It should be noted that the above-mentioned analysis and processing of organizational structure and functional division information involves stripping away non-direct business content such as organizational hierarchy, departmental affiliation, job positions, reporting relationships, and management rules. The core purpose is to extract the specific responsibilities, work content, areas of responsibility, business jurisdiction, work nodes, and scope of authority corresponding to each department and position, and to summarize the organization's business information.

[0037] S22, The unit business information is processed using a hierarchical classification method to obtain an initial unit business classification system; It should be noted that a top-down, progressively subdivided hierarchical classification method is used to organize and structure unit business information in layers, constructing a tree-like hierarchical structure. This achieves the orderly classification of business items, forming an initial unit business classification system, and providing a unified classification standard for subsequent tag construction, data annotation, and model business adaptation. Generally, a three-layer division is adopted, specifically: The first level is divided according to the core functional areas of the unit, such as administrative management, business development, Party building, logistical support, and external cooperation. The second level is divided according to internal departments, such as the office, business department, human resources department, finance department, project team, etc. The third level is divided according to specific business matters, such as meeting organization, document drafting, project implementation, activity development, training organization, inspection and supervision, etc.

[0038] S23, The initial unit business classification system is modified to obtain the unit business classification system; It should be noted that the above-mentioned revision process involves verifying, merging, splitting, supplementing, and standardizing the initial unit business classification system to form a logically rigorous, clearly defined, comprehensive, and consistent system. Verification primarily involves eliminating duplicate, overlapping, and logically contradictory classification items; merging primarily involves merging semantically similar and ambiguous business categories; splitting primarily involves splitting broad categories with multiple independent subcategories; supplementing primarily involves adding missing core businesses, special tasks, and major matters; and standardization primarily involves unifying classification names, definitions, and hierarchical naming rules. An example of a municipal government department's business classification system is as follows: Primary categories: Administrative management, business performance, Party building and integrity, project management, meetings and activities, and external cooperation; Secondary classification: Administrative management: document management, personnel management, financial management, and logistical support; Business duties: policy formulation, supervision and inspection, service processing, and data statistics; Party building and anti-corruption: Party building activities, anti-corruption efforts, learning and education, and ideology; Project management: project initiation, project implementation, project acceptance, and results dissemination; Meetings and activities: Party group meetings, office meetings, special meetings, major events; External collaboration: liaison with higher authorities, inter-departmental coordination, grassroots research, and external exchanges; Third-level classification: Administrative Management - Document Management: Issuance, receipt, approval, correspondence, and briefings; Administrative Management - Human Resources Management: Recruitment, training, performance evaluation, appointment and dismissal, personnel adjustments; Business Performance - Supervision and Inspection: Special supervision, routine inspection, rectification supervision, and on-site verification; Conferences and Events - Major Events: Anniversary Celebrations, Unveiling Ceremony, Awards Ceremony, Signing Ceremony.

[0039] S24. Based on the unit business classification system, process the unit business information to obtain a unit business classification tag library.

[0040] It should be noted that, based on the revised unit business classification system, the unit's business information is sorted, refined, and standardized. This involves extracting and standardizing various business keywords, professional terms, business categories, work items, and function-related vocabulary. Mapping, format unification, deduplication, and structured coding are then performed according to the classification system hierarchy, ultimately forming a unique, standardized, and fully adaptable unit business classification tag library that can be searched and matched. An example of a municipal government department's business classification tag library is shown in the table below: It should be noted that when an organization undergoes organizational restructuring, changes in functional division, addition of core businesses, or adjustments to policy requirements, the processing steps S21 to S24 will be repeatedly executed to update the aforementioned organization's business category tag library.

[0041] As can be seen, the unit business record set generation method based on a large model described in this embodiment of the invention, through a closed-loop process of "parsing business information—hierarchically constructing initial classifications—correcting and optimizing the classification system—building a dedicated tag library," extracts and constructs a hierarchical, standardized, and comprehensive business classification system and tag library adapted to the unit's actual situation from the unit's organizational structure and functional division information. This effectively solves the problems of mismatch between general classification systems and unit-specific business, inconsistent definitions, semantic inconsistencies, and incomplete coverage, providing a unified semantic standard and classification basis for subsequent structured processing of business materials, accurate annotation, and model business adaptation training.

[0042] In another optional embodiment, the step of processing the unit business material set based on the unit business classification system and the unit business classification label library to obtain a training dataset and a validation dataset includes: S31, Establish a unit business material index library based on the aforementioned unit business material set; It should be noted that, based on the unit's business materials collection, unified cataloging, classification, numbering, and information registration are carried out according to the dimensions of year, department, business type, material category, storage path, and source channel. Each business material is assigned a unique index identifier, recording index metadata such as material name, creation time, department, business affiliation, media type, storage location, and material summary, thus constructing a structured, quickly searchable, and locatable unit business materials index library. This unit business materials index library is used to achieve orderly management, rapid retrieval, and accurate location of massive amounts of business materials, providing index support for subsequent batch data extraction and standardized processing. The above-mentioned unit business materials index library can be built using file system directory hierarchical indexing technology, relational database structured database construction technology, or full-text search engine index construction technology; this embodiment of the invention does not limit the specific methods used.

[0043] S32, Based on the unit business material index library, perform data extraction processing on the unit business material set to obtain the unit business dataset; It should be noted that, relying on the unit's business materials index library, precise retrieval and batch retrieval of the unit's business materials sets that have been fully digitized are performed by year, department, business category, and storage path. Professional data extraction and standardized processing are then carried out on the retrieved digitized business materials, uniformly parsing various electronic documents, minutes, briefings, ledgers, image metadata, and work dynamic text content to extract core business information such as effective business text, event descriptions, and attribute metadata. Simultaneously, format unification, redundant field removal, duplicate data deduplication, invalid noise filtering, content normalization, and compliance verification are performed to form a standardized, formatted, and complete unit business dataset, providing a standardized basic data source for subsequent information annotation, dataset partitioning, and model training.

[0044] S33, Based on the unit business classification system and the unit business classification label library, the unit business dataset is annotated with relevant information of the record set to obtain the unit business labeled dataset; It should be noted that the above information annotation process uses the unit business classification system as the hierarchical classification framework and the unit business classification tag library as the unified semantic annotation standard to perform specialized information annotation processing on the unit business dataset for the generation of unit business chronicles. Combining actual business scenarios, time information, entity information, event information, information source, and business domain tags are identified and labeled sentence by sentence and paragraph by paragraph from the business data. At the same time, the annotated content is associated, mapped, and matched with the hierarchical categories of the business classification system and the standard tags of the business classification tag library. In the annotation process, the annotation criteria are unified, the element boundaries are standardized, and the annotation ambiguity is eliminated, forming a unit business annotated dataset that is complete in elements, well-classified, and has unified tags, and is suitable for the unit business chronicle generation task. This provides a standardized and reliable annotation sample foundation for subsequent dataset partitioning and model training and verification.

[0045] It should be noted that when annotating time information, the time of event occurrence or the time of material formation should be accurate to a specific day and hour. If a time range exists, such as a project cycle of 202X.01-202X.12, the complete time range should be annotated. When annotating entity information, priority should be given to annotating core entities related to the business chronicle, including unit name, department name, core personnel, partner units, project name, honor name, key locations, etc. The entity type should be clearly stated when annotating, such as "Project Name - XX Digital Transformation Project" or "Honor Name - XX Industry Innovation Award". When annotating event information, the specific type of event should be clearly stated, such as project launch, meeting held, honor received, cooperation agreement signed, policy implementation, results released, etc. The key status of the event should also be annotated, such as project launch or meeting held. When annotating the source of information, the name of the original material of the sample, storage location, and acquisition channel should be clearly stated. When annotating business domain tags, accurate business category tags should be annotated according to the unit's business classification system and unit business classification tag library.

[0046] S34, Select and divide the unit business annotation dataset to obtain the training dataset and the validation dataset.

[0047] It should be noted that the specific process of selecting and dividing the unit business annotation dataset is as follows: First, the unit business annotation dataset is screened to remove low-quality samples that are incomplete, have incorrect annotations, contain redundant information, or do not meet the requirements of the unit business chronicle generation task, while retaining effective annotation data that is well-labeled, complete in elements, and has balanced business coverage. Then, according to the preset division ratio, following the principles of cross-year balance, uniform distribution across business areas, and unbiased sample types, the effective annotation dataset is split and divided using stratified random sampling. This generates training datasets for model iteration training and parameter optimization fitting, as well as validation datasets for independent effect evaluation, generalization ability verification, and model performance verification. This provides standard sample data that is evenly distributed, reliable in quality, and has broad business coverage for subsequent full-process training and validation of data analysis models.

[0048] It should be noted that in the above validation dataset, regular samples account for 70%, difficult samples account for 20%, and noisy samples account for 10%.

[0049] As can be seen, the unit business record set generation method based on a large model described in the embodiments of the present invention completes the full-link standardization process of unit business materials from original electronic data to usable model sample data by constructing a unit business material index library, extracting business data, standardizing and annotating multi-dimensional full-element data, and filtering and dividing the labeled dataset in a complete closed loop. This lays a solid foundation of high-quality data for the subsequent training convergence, feature learning, and business adaptation capabilities of the data analysis model.

[0050] In another optional embodiment, the step of training and validating a preset data analysis model using the training dataset and the validation dataset to obtain a unit business record generation model includes: S41, using the training dataset, train the time information extraction sub-model, the entity information extraction sub-model, the event information extraction sub-model, and the image association and information source extraction sub-model respectively, to obtain the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model; It should be noted that this step utilizes the constructed training dataset to conduct independent training on the time information extraction sub-model, entity information extraction sub-model, event information extraction sub-model, and image association and information source extraction sub-model, respectively. Through supervised learning, the model parameters of the feature extraction layer and output layer within each sub-model are iteratively optimized, thereby improving the ability of each sub-model to identify, extract, and analyze the corresponding elements. After training, the optimized sub-models corresponding to each sub-model are obtained in sequence.

[0051] S42, the Transformer is used to fuse and collaboratively optimize the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model to obtain a multi-task collaborative data analysis model; It should be noted that Transformer possesses powerful global context modeling capabilities, multi-head attention mechanisms, and cross-semantic association modeling advantages. It can accurately capture the implicit semantic dependencies and logical connections between four types of tasks: time information extraction, entity information extraction, event information extraction, image association, and information source extraction. It can achieve global interaction and weighted fusion of multi-task features. Compared with single convolutional and recurrent networks, it is more suitable for multi-task parallel collaborative optimization scenarios. Therefore, Transformer is used to perform deep feature fusion, task association constraints, and global collaborative parameter tuning on four trained and optimized sub-models. This breaks down the information barriers of independent operation of each sub-model, realizes multi-task feature complementarity, semantic linkage, and joint parameter convergence, and integrates and constructs a multi-task collaborative data analysis model.

[0052] S43, using the unit business classification tag library, train the business domain matching capability of the multi-task collaborative data analysis model to obtain the first unit business record generation model; S44, perform text data augmentation training on the first unit business record generation model to obtain the second unit business record generation model; It should be noted that the above-mentioned text data augmentation training uses text augmentation training samples to iteratively train the first unit business chronicle generation model, which effectively improves the model's ability to generalize and understand business materials with different writing styles, sentence structures, and expression habits, alleviates the risk of overfitting caused by the limited real labeled samples, and further strengthens the model's basic representation ability. After the training converges, the second unit business chronicle generation model is obtained. Text augmentation training samples were obtained using the following methods: The training samples in the training dataset were expanded using methods such as synonym replacement, sentence transformation, and text fragment recombination. The replacement scope was strictly controlled for professional terms within the unit's business, replacing only keywords related to the professional terms to avoid semantic bias caused by replacing the professional terms themselves. For short text summaries of less than 100 characters, corresponding contextual semantic extensions were added to enrich the semantic information of the short text and avoid the model's over-reliance on local expressions, which could lead to element extraction bias. Noisy text samples were added, including business texts containing typos, colloquial expressions, and disordered word order. These samples accounted for approximately 10% of the augmented samples. The noise filtering ability of the model was trained using these samples, improving the model's adaptability and parsing ability for non-standard business texts.

[0053] S45, perform adversarial optimization training on the second unit business record generation model to obtain the third unit business record generation model; It should be noted that the specific implementation method for adversarial optimization training of the second unit business chronicle generation model is as follows: A fast gradient symbolic method adversarial perturbation technique is introduced, adding a small perturbation to the text word embedding layer of the model. This simulates scenarios where business texts have slight semantic interference and expression deviations, allowing the training model to maintain the stability of key information extraction even under perturbation. The adversarial training weight is set to 0.15, and the adversarial training loss is superimposed with the regular training loss as the total loss function for iterative model training, guiding model parameter optimization. Four rounds of iterative training are conducted according to this setting. After each round, the model performance decay rate under adversarial scenarios is verified, strictly controlling the decay rate to not exceed 5%, ensuring that the model can still maintain high element extraction accuracy and robustness under adversarial scenarios. After the above adversarial optimization training converges, the third unit business chronicle generation model is obtained.

[0054] S46, perform cross-format text adaptation training on the third unit business record generation model to obtain the fourth unit business record generation model. It should be noted that the specific implementation method for cross-format text adaptation training of the third unit business record generation model is as follows: Select text samples of different formats in the unit's business, specifically including common business text formats such as meeting minutes, press releases, contracts and agreements, and work reports. Group them according to text format type, and conduct adaptation training for different format groups separately. Combining the unique structural characteristics of each format text, focus on optimizing the local attention window size of the model's feature extraction layer. For example, for the typical structure of meeting minutes ("time-issue-resolution"), the core structure of contracts and agreements ("terms-responsible party-amount"), and the exclusive structure of other format texts, adjust the attention window parameters in a targeted manner to enhance the model's ability to capture structured information of different format texts. This enables the model to accurately adapt to the element distribution patterns of various format business texts, improve the unified parsing and key information extraction capabilities of cross-format texts, and after the above cross-format text adaptation training is completed, the fourth unit business record generation model is obtained.

[0055] S47, Using the verification dataset and preset multi-dimensional verification indicators, a verification experiment is conducted on the fourth unit business record generation model to obtain the compliance verification result; When the compliance verification result is yes, the fourth unit business record set generation model is determined to be a unit business record set generation model; If the compliance verification result is negative, repeat steps S43 to S47 until the compliance verification result is positive.

[0056] It should be noted that the specific settings for the above multi-dimensional verification indicators are as follows: Overall F1 score for key information extraction ≥ 93%, with each subtask's F1 score ≥ 90%; Business domain matching accuracy calculated by hierarchical category average ≥ 92%; Performance degradation rate in generalization scenarios ≤ 5%; Single text processing speed ≤ 0.5 seconds / text; The F1 value is the harmonic mean of recognition accuracy (Acc) and recall (Rec), comprehensively reflecting the model's accuracy and completeness. The expressions for F1, recognition accuracy (Acc), and recall (Rec) are as follows: It should be noted that once all validation metrics of the model reach the preset thresholds and the validation results meet the standards, all model parameters are immediately solidified, and a complete model file is generated. This file includes the model weight file, configuration file, and label mapping table. At the same time, key parameters during the training process are recorded in detail, including the learning rate, loss function weights, iteration rounds, etc., forming a complete model training report. The solidified model unit business classification label library and preprocessing database are seamlessly integrated, and the final output is a dedicated analysis model that can be directly used for generating the target year's unit business chronicle. If the model's validation results are negative, training steps S43 to S46 are repeated until the model meets the multi-dimensional validation metrics, resulting in a dedicated analysis model that can be directly used for generating the target year's unit business chronicle. It should be noted that in the subsequent practical application of the model, a normalized iterative update mechanism should be established, and the actual application accuracy of the model should be statistically analyzed every quarter. When the actual application accuracy is more than 5% lower than the relevant verification indicators, the model iterative training process should be triggered. Based on the newly added business samples and changes in the unit's business, incremental training should be carried out according to the training process of this embodiment to update the relevant parameters of the model, so as to ensure that the model always adapts to the needs of the unit's business development and maintains a stable and accurate ability to generate unit business chronicles.

[0057] As can be seen, the unit business chronicle generation method described in this embodiment of the invention completes the entire process optimization of the data analysis model through a hierarchical progressive training architecture: first, four types of information extraction sub-models are trained independently; then, a multi-task collaborative model is constructed using Transformer fusion; and through business domain matching, text data augmentation, fast gradient symbolic adversarial optimization, and cross-format adaptation training, the model's adaptability, robustness, and parsing ability are enhanced. Verification is performed through multi-dimensional quantitative indicators, and once the standards are met, the parameters are solidified and connected to relevant databases. Simultaneously, a quarterly iteration mechanism is established, ensuring the accuracy, business relevance, and long-term adaptability of the unit business chronicle generation model, providing stable core support for the intelligent generation of high-quality unit business chronicles.

[0058] In another optional embodiment, training the event information extraction sub-model using the training dataset to obtain a corresponding optimized event information extraction sub-model includes: S411, Construct an event type and trigger word mapping library based on the unit's business material set; It should be noted that the method for constructing the above-mentioned event type and trigger word mapping library is as follows: Based on the full set of business texts covered by the unit's business materials collection, including official documents, minutes, work updates, project materials, and activity records, the unit's full set of standard event types are sorted and summarized, such as meeting convening, project initiation, research and supervision, commendation and awards, cooperation signing, and policy implementation; then, the unit's characteristic trigger words, business keywords, and behavioral feature words corresponding to each type of event are mined, and a one-to-one correspondence mapping relationship between event types and trigger words is established. After sorting, classifying, deduplicating, and unifying, a structured event type and trigger word mapping library adapted to the unit's business scenarios is constructed, providing a standard thesaurus basis for subsequent accurate annotation of event elements.

[0059] S412, using the event type and trigger word mapping library, perform event trigger word annotation, event type annotation and event status annotation processing on the training dataset to obtain an event annotation sample set; It should be noted that the above annotation process is based on the event type and trigger word mapping library, and performs standardized three-element annotation on business text in the training dataset: accurately locate and annotate event trigger words in the text, match and annotate the corresponding standard event type, and annotate event status such as event initiation, progress, and completion; unify annotation standards, standardize element boundaries, eliminate semantic ambiguity, and complete the annotation of all samples in batches to form an event annotation sample set with complete elements, standardized annotation, and unified standards.

[0060] S413, Perform trigger word context enhancement and sample balancing on the event-labeled sample set to obtain the event training sample set; It should be noted that the above-mentioned trigger word context enhancement processing expands the semantic fragments of the context before and after the event trigger word, enriches the semantic features associated with the trigger word, and avoids misjudgment caused by the model relying solely on isolated keywords; the above-mentioned sample balancing processing performs quantity supplementation, sampling balancing, and deduplication cleaning of various event samples, dilutes the proportion of high-frequency event samples, supplements the number of rare event samples, alleviates the problem of unbalanced category distribution, and obtains a balanced event training sample set with complete semantic features after processing.

[0061] S414, using the event type training weight calculation model and the event training sample set, the event information extraction sub-model is jointly fitted and trained with event type and event state to obtain the event information extraction sub-model to be optimized; the expression of the event type training weight calculation model is: In the formula, The normalized training weights for the k-th class of event samples take values ​​in the range (0,1). Let be the normalized frequency of the k-th type of event, with a value range of (0,1]. This is a preset frequency enhancement factor, with a value range of [0.3, 2.0]. Let be the sample difficulty coefficient for the k-th type of event. The initial prediction confidence of the event information extraction sub-model for the k-th type of event sample is given, with a value range of [0,1). Let be the rarity coefficient of the k-th type of event. Let k be the number of samples of the k-th type of event. The total number of samples in the training sample set for the event. The preset rarity compensation factor has a value range of [0.3, 2.0]; K is the total number of event types in the event training sample set; It should be noted that the sum of the normalized training weights of all types of event samples in the above event training sample set is 1; the normalized frequency of the above k-th type of event... The frequency enhancement factor is obtained by normalizing the proportion of the number of event samples of type k to the event training sample set; the preset frequency enhancement factor is as follows. This is used to adjust the intensity of the influence of the frequency dimension on the weights; the aforementioned preset rarity compensation factor This is used to adjust the strength of the compensation for the weights based on the rarity dimension. It should be noted that in the above-mentioned fitting and joint training process, the event type-based training weight calculation model dynamically assigns training weights to different types of event samples, realizing differentiated reinforcement learning for high-frequency events, difficult-to-identify events, and rare events. The event type-based training weight calculation model dynamically calculates training weights from three dimensions: event frequency, identification difficulty, and sample rarity. In the frequency dimension, high-frequency events with a high sample ratio are assigned higher base weights to ensure the stability of regular event identification; in the difficulty dimension, weights are calculated based on... It is determined that the lower the initial prediction confidence, the higher the sample identification difficulty, and the greater the corresponding difficulty coefficient. This increases the training weight of difficult examples with ambiguous boundaries and easy confusion, strengthening the model's ability to identify difficult examples. The rarity dimension increases with the number of event samples, automatically raising the compensation weight for small samples and low-probability major events, effectively avoiding missed detections of rare key events. Through weighted joint fitting training, the event type and event state features converge synchronously, enabling the model to learn various event features in a balanced way, simultaneously improving the accuracy of event type classification and the precision of event state determination, resulting in the sub-model for extracting event information to be optimized.

[0062] It should be noted that the initial prediction confidence of the event information extraction sub-model for the k-th type of event sample is the probability value output by the event information extraction sub-model when performing preliminary type identification of the k-th type of event sample before formal training and after completing basic feature extraction. The model first encodes the text of the k-th type of event into a feature vector, and then outputs the confidence probability that the sample belongs to the k-th type of event through the built-in preliminary classification branch. The value range is [0,1). The closer the value is to 1, the higher the accuracy of the model's preliminary identification and the lower the difficulty of sample identification. The closer the value is to 0, the greater the deviation of the model's preliminary identification and the more difficult the sample is.

[0063] S415, perform time and event state correlation verification and feature constraint optimization on the event information extraction sub-model to be optimized to obtain the event information extraction optimization sub-model.

[0064] It should be noted that the above-mentioned time and event status association verification processing is based on the preset logical constraint relationship between the event occurrence time, time period range and event status. It verifies and corrects the identification results of mismatch and logical conflict between time elements and event status. For example, the event of a project under construction should not be matched with the completion time node, and the event of annual work summary should not be identified as the start status at the beginning of the year. The logical association verification corrects such time sequence and status mismatch problems. The above-mentioned feature constraint optimization processing applies business semantic constraints and event boundary constraints to the model feature extraction layer, solidifies effective business features and suppresses irrelevant interference features. After the dual optimization of association verification and feature constraints, the accuracy of model event classification, status determination and time sequence association recognition is further improved, and finally the optimized sub-model of event information extraction is obtained.

[0065] As can be seen, the unit business record generation method based on a large model described in the embodiments of the present invention achieves multi-dimensional standardized labeling of events by constructing an event type and trigger word mapping library; optimizes sample quality through context enhancement and sample balancing; dynamically allocates training weights through a multi-dimensional weight model to differentiate high-frequency, difficult-to-identify, and rare events; and combines time state correlation verification and feature constraint optimization to effectively avoid missing rare events and confusing similar events, thereby significantly improving the accuracy of event classification and state determination of the model.

[0066] In another optional embodiment, the step of training the image association and information source extraction sub-model using the training dataset to obtain the corresponding optimized image association and information source extraction sub-model includes: A411, perform text and metadata association annotation and credibility filtering on the training dataset to obtain an association annotation sample set; The metadata is supplementary text data that corresponds one-to-one with the unit business image data in the training dataset, used to describe the basic attributes of the image, trace the source of information, identify the storage location and related business scenarios, including image identification, storage path, information source channel, acquisition time, department to which it belongs and credibility rating; It should be noted that the above credibility rating is assigned to the credibility of the unit business image data in the above training dataset according to the source of information. For example, the credibility of unit business image data from superior documents is 10, the credibility of unit business image data from internal official documents is 9, the credibility of unit business image data from official media reports is 8, and the credibility of unit business image data from ordinary online news is 6. It should be noted that the above text and metadata association annotation uses the image identifier as the unique association primary key, and binds and associates the unit business image data, the corresponding image metadata, and the business event text data one by one, establishing a one-to-one correspondence mapping relationship between "business text - image data - image metadata", and annotating and defining the business affiliation, scene association, and source association between each data, so as to achieve accurate connection and matching of multi-source data; the above credibility screening process is based on the credibility rating of metadata, and removes metadata with a credibility rating of less than 8 and its corresponding unit business image data and business text.

[0067] A412, based on the aforementioned associated labeled sample set, positive and negative samples are constructed and material type grouping is performed to obtain a multi-scene associated training sample set; It should be noted that the above-mentioned positive and negative sample construction and material type grouping are based on the associated labeled sample set for sample construction and category regularization: on the one hand, positive samples that correctly match images and business texts and irrelevant mismatched negative samples are constructed to enrich the model's association discrimination learning scenarios; on the other hand, the samples are grouped and classified according to business material types such as official documents, meeting materials, project archives, and news images to distinguish the image association features and source attribute patterns under different business scenarios. After positive and negative sample balancing and grouping and regularization by type, a multi-scenario association training sample set covering multiple business scenarios is obtained.

[0068] A413, using the multi-scene association training sample set, the image association and information source extraction sub-model is trained by cross-modal feature alignment fitting to obtain the image association and information source extraction sub-model to be optimized; It should be noted that the aforementioned cross-modal feature alignment and fitting training is a model-specific training exercise conducted for three different modalities of data: business text, image data, and image metadata. The text modality carries business semantics and event information, the image modality carries scene and image features, and the metadata modality carries attribute information such as source, time, department, and credibility. The original feature dimensions and representation spaces of the different modalities are not uniform. During training, feature encoding is first performed on the text, image, and metadata separately. Then, modal mapping transformation is used to map the three types of features to the same feature space, completing dimensionality normalization and semantic alignment, and eliminating heterogeneous differences between modalities. Subsequently, the sample real association annotation is used as a supervision signal to perform joint loss fitting training, allowing the model to learn the inherent association matching rules between text content, image scene, and metadata attributes, achieving multi-modal feature complementarity and accurate identification of association relationships, thereby completing cross-modal feature alignment and joint fitting convergence. After training, the sub-model for image association and information source extraction to be optimized is obtained.

[0069] It should be noted that the above-mentioned cross-modal feature alignment fitting training can be performed with the goal of minimizing the difficulty-aware dynamic temperature contrast loss function, so as to obtain the sub-model of image association and information source extraction to be optimized. The expression for the difficulty-aware dynamic temperature contrast loss function is: In the formula, The dynamic temperature contrast loss value is the difficulty-aware loss value, which is the optimization target for training the image association and information source extraction sub-model, and its value ranges from [0, +]. During training, this value is continuously reduced through backpropagation, enabling the model to learn the correct cross-modal correlation features; The sample difficulty weight is the i-th training sample in the multi-scene association training sample set, and its value ranges from [0,1]. The initial matching confidence score of the image association and information source extraction sub-model for the i-th training sample is obtained from the model's initial matching probability output of text and image metadata features. Its value ranges from [0,1]. The higher the sample matching difficulty, the lower the initial confidence score, and the greater the difficulty weight. This is used to increase the proportion of difficult samples in the loss calculation and strengthen the model's learning ability for ambiguous and easily confused samples. sim(,) is the cosine similarity function; s is the positive sample similarity, and is the text feature vector. Matching image metadata feature vectors The cosine similarity is calculated, with values ​​ranging from [-1, 1]. A value closer to 1 indicates a higher semantic correlation between the text and the matched image metadata. Negative sample similarity, text feature vector. Feature vectors of mismatched image metadata The cosine similarity is taken in the range of [-1, 1]. The closer the value is to -1, the greater the semantic difference between the text and the metadata of the mismatched image. The dynamic temperature coefficient is determined dynamically by the initial temperature and sample difficulty. It is used to control the sharpness of the similarity distribution in the contrast loss. The lower the temperature, the more obvious the distinction between positive and negative samples; the higher the temperature, the smoother the distribution. The preset initial temperature hyperparameter has a value range of (0.8, 1.2), with a preferred value of 1.0 to ensure a smooth and moderate initial distribution.

[0070] It should be noted that the training method using the aforementioned difficulty-aware dynamic temperature contrast loss function is as follows: First, the text and image metadata are encoded into feature vectors respectively, and the similarity between positive and negative samples, the sample difficulty weight, and the dynamic temperature coefficient are calculated. Then, the loss value of the current sample is calculated by substituting them into the loss function, and the model parameters are updated through the backpropagation algorithm. With the goal of minimizing the loss value, the feature distance between the text and the matching image metadata is continuously reduced, while the feature distance between the text and the unmatched image metadata is increased. At the same time, the learning of difficult examples is strengthened by the difficulty weight, and the difficulty is distinguished by the dynamic temperature to adapt different samples. Finally, the model has the ability to accurately associate text and image metadata across modalities, and the sub-model of image association and information source extraction to be optimized is obtained.

[0071] It should be noted that the initial matching confidence of the image association and information source extraction sub-model for the i-th training sample is the matching probability value output by the built-in preliminary matching branch before the formal training of the image association and information source extraction sub-model. The value range is [0,1], which is used to measure the sample matching difficulty. The closer the value is to 1, the simpler the sample is, and the closer it is to 0, the more difficult the sample is.

[0072] A414, perform association matching accuracy calibration processing on the sub-model of image association and information source extraction to be optimized, and obtain the optimized sub-model of image association and information source extraction.

[0073] It should be noted that the above-mentioned correlation matching accuracy calibration process is based on the standard matching relationship of the correlation annotation sample set. It compares the current output of the business text and image correlation matching results of the model, and counts the mismatch, missing match, and multiple match samples. For the incorrectly matched samples, it reversely corrects the model's cross-modal feature correlation threshold and matching discrimination boundary, fine-tunes the feature similarity calculation rules, and constrains the model to strictly follow the business scenario, the department to which it belongs, and the event logic for correlation matching. It eliminates invalid correlations, corrects incorrect correlations, and comprehensively improves the accuracy, rigor, and compliance of correlation matching between text and images.

[0074] As can be seen, the unit business record set generation method based on a large model described in the embodiments of the present invention ensures the reliability and regularity of training samples through image identification association annotation and credibility hierarchical screening; constructs positive and negative samples and groups them according to material type to form a multi-scenario training sample set; adopts cross-modal feature alignment fitting training, combines difficulty-aware dynamic temperature contrast loss function to enhance difficult example learning, and then corrects the deviation through association matching accuracy calibration, effectively improving the model's cross-modal association matching accuracy and credibility discrimination ability.

[0075] In another optional embodiment, training the time information extraction sub-model using the training dataset to obtain a corresponding optimized time information extraction sub-model includes: B411, Perform time location annotation and format normalization processing on the training dataset to obtain a time-annotated sample set; It should be noted that the above-mentioned time location annotation and format standardization process involves annotating the time location of each sentence in the business text of the training dataset, accurately locating the character positions and boundary ranges of time elements such as explicit time, implicit time, and time period range in the text; at the same time, time format standardization is performed to unify the scattered and inconsistent time expressions into standardized formats such as year, month, day, time period, and relative time sequence, unify the annotation criteria and format standards, remove redundant expressions, and obtain a time annotation sample set with accurate annotation positions and uniform and standardized formats.

[0076] B412, perform temporal feature enhancement and format compliance screening on the time-marked sample set to obtain a time-enhanced sample set; It should be noted that the above-mentioned temporal feature enhancement and format compliance screening processes are based on expanding the semantics of contextual temporal association around the labeled time elements, strengthening the expression of implicit temporal features such as the logical sequence of events, the order of events, and the constraints of time periods. Then, format compliance screening is performed to remove unqualified samples with time annotation errors, seriously non-standard formats, and conflicting temporal logic, and retain high-quality samples with complete temporal features, compliant formats, and logical consistency. After feature enhancement and compliance screening, a time-enhanced sample set with complete temporal information and controllable sample quality is obtained.

[0077] B413, using the time-enhanced sample set, the time information extraction sub-model is trained by fitting time-series features to obtain the time extraction sub-model to be optimized; It should be noted that the above-mentioned time-series feature fitting training adds an independent time-specific recognition branch to the output layer of the time information extraction sub-model. A bidirectional LSTM+CRF structure is used to enhance time-series feature capture, adapting to the sub-model's own feature extraction needs and improving the ability to perceive sequences and identify boundaries for continuous time text and irregular time expressions. During training, the time label weight is set to 1.2, higher than that of ordinary entity labels, to strengthen the learning priority of time elements. A learning rate is configured accordingly. The training process involves six rounds of iteration, with each round verifying the accuracy of time extraction and real-time monitoring of the model's convergence status. Through supervised learning, the model fits time element features, identifies boundaries, and tracks temporal logic patterns. After training convergence, a sub-model for time extraction to be optimized is obtained.

[0078] B414, perform time identification weight calibration and basic parameter solidification on the time extraction sub-model to be optimized to obtain the time information extraction optimization sub-model.

[0079] It should be noted that the above-mentioned time recognition weight calibration and basic parameter solidification process first involves calibrating the time recognition weights, fine-tuning the model's feature recognition weights for different types of time expressions, long and short time periods, and relative time sequence scenarios, enhancing the recognition sensitivity of implicit time and ambiguous time periods, and correcting omissions and misidentifications. After the time extraction performance meets the standards, the basic parameter solidification process is performed to freeze the core basic parameters of the time-specific recognition branch, avoiding parameter perturbations during subsequent training, and finally outputting an optimized sub-model for time information extraction that is accurate in recognition, has reliable temporal logic, and stable parameters.

[0080] As can be seen, the unit business chronicle generation method described in this embodiment of the invention constructs high-quality training samples through processing steps such as time location annotation, format standardization, temporal feature enhancement, and compliance screening; it adopts a bidirectional LSTM+CRF structure, adds a dedicated time recognition branch, and strengthens the learning priority of time elements by combining reasonable hyperparameter settings and weight allocation; through optimizations such as weight calibration and parameter solidification, it effectively improves the accuracy and stability of time information extraction, avoids recognition deviation and parameter disturbance, solves problems such as non-standard time extraction and inaccurate recognition, ensures efficient and reliable time information extraction, and provides solid time extraction support for the subsequent unit business chronicle to be chronologically regular and accurate in content.

[0081] In another optional embodiment, training the entity information extraction sub-model using the training dataset to obtain a corresponding optimized entity information extraction sub-model includes: C411, Perform entity boundary annotation and type classification on the training dataset to obtain an entity annotation sample set; It should be noted that the above entity boundary annotation and type classification processing involves annotating the entity boundaries of each sentence of the unit business text in the training dataset, accurately defining the start and end positions of various business entities, and clarifying the boundary range of entity characters. At the same time, the annotated entities are classified into categories such as organizational entities, personnel entities, departmental entities, project entities, location entities, and honorary entities according to business attributes. The entity annotation rules and type classification standards are unified, invalid and redundant annotations are removed, and after regularization, a sample set of entity annotations with clear entity boundaries and standardized type classification is obtained.

[0082] C412, perform entity frequency grouping and boundary fuzzy sample enhancement processing on the entity annotation sample set to obtain the entity training sample set; It should be noted that the specific process of the above-mentioned entity frequency grouping and boundary ambiguity sample enhancement is as follows: First, core business entities that appear frequently in the entity annotation sample set are prioritized for grouping, such as unit names and project names, to highlight the training proportion of key entities. Then, low-frequency entities are also considered, such as key locations and honorary names, in line with the training logic of "prioritizing training high-frequency entities and then training low-frequency entities". Then, feature enhancement is performed on difficult-to-distinguish entity samples with ambiguous boundaries, easy confusion, and similar contextual semantics, such as similar project names, to expand the semantics of the associated context, highlight the entity distinguishing features, implement the difficult example mining strategy, and lay the foundation for subsequent individual enhancement training. After high-frequency entity grouping, low-frequency entity consideration, and boundary ambiguity sample feature enhancement, samples with qualified quality and high feature discrimination are selected and retained to form an entity training sample set suitable for multi-category entity learning.

[0083] C413, using the entity training sample set, perform multi-class feature fitting training on the entity information extraction sub-model to obtain the entity information extraction sub-model to be optimized; It should be noted that the above multi-category feature fitting training strictly follows the principle of "grouping training by entity type": high-frequency entities are trained first, and low-frequency entities are trained only after the high-frequency entity recognition accuracy reaches the standard, ensuring the priority of core entity recognition. During training, a difficult example mining strategy is adopted, with separate reinforcement training for samples with ambiguous entity boundaries and easily confused types, controlling the proportion of difficult examples in each round to be no less than 20%, thus strengthening the model's ability to recognize difficult examples. For core personnel entities, "name + position" joint recognition training is supplemented to avoid the problem of only recognizing names and not positions, improving the completeness of personnel entity extraction. Relying on the sub-model's own independent feature extraction layer and output layer, the semantic features, contextual expression rules, and boundary representation differences of different types of business entities are learned, fitting the recognition and classification logic of multi-category entities. Through supervised training, the model's feature extraction ability and entity classification ability are continuously optimized. After training convergence, a sub-model for entity information extraction with preliminary entity recognition and classification capabilities is obtained.

[0084] C414, perform entity type differentiation calibration and model parameter convergence processing on the entity information extraction sub-model to be optimized to obtain the entity information extraction optimization sub-model.

[0085] It should be noted that the above entity type differentiation calibration and model parameter convergence processing first involves entity type differentiation calibration, which fine-tunes the model's feature discrimination weights for similar entities with high similarity and easy confusion, optimizes the differentiation boundaries of different entity types, corrects entity misclassification, omission, and category judgment bias, and further consolidates the training effect of difficult sample samples. Then, model parameter convergence processing is performed. After the model's entity recognition accuracy and the accuracy of the joint recognition of "name + position" of personnel entities reach the target, the convergence optimization and stabilization tuning of the model's core parameters are completed to prevent parameter fluctuations in subsequent training. Finally, an optimized sub-model for entity information extraction is obtained, which has accurate entity boundary localization, clear type differentiation, stable recognition performance, and complete extraction of personnel entities.

[0086] As can be seen, the unit business record set generation method based on a large model described in the embodiments of the present invention constructs a standardized entity annotation sample set through entity boundary annotation and type classification; high-quality training samples are formed by prioritizing high-frequency entities and strengthening samples with blurred boundaries; during training, the group training principle is followed, prioritizing the training of core high-frequency entities, combining hard example mining and strengthening with joint recognition of personnel entities "name + position", and then through type differentiation calibration and parameter convergence optimization, effectively improving the accuracy, type differentiation and completeness of entity recognition, avoiding erroneous recognition, and providing reliable support for the accurate extraction of core entities in the unit business record set.

[0087] In another optional embodiment, the step of using Transformer to fuse and collaboratively optimize the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model to obtain a multi-task collaborative data analysis model includes: S421, the feature extraction layers of the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model are integrated using Transformer to obtain a feature sharing layer; It should be noted that the above integration of the feature extraction layers of the four optimized sub-models is based on the Transformer network architecture, and integrates the built-in feature extraction layers of the time information extraction optimized sub-model, entity information extraction optimized sub-model, event information extraction optimized sub-model, and image association and information source extraction optimized sub-model, respectively. The network structure, semantic representation parameters, and text feature mapping rules of the original independent feature extraction layers of the four sub-models are uniformly integrated, weighted and aligned, eliminating the parameter redundancy and feature fragmentation problems caused by the separate feature extraction of each sub-model. Through the global semantic modeling capability of Transformer, the global features and local key features of the text are uniformly learned to form a feature sharing layer that can provide general semantic support for the four sub-tasks at the same time, realizing the reuse of low-level features and unified representation for multiple tasks.

[0088] S422, taking the feature sharing layer as the core, the output layer of the time information extraction optimization sub-model, the output layer of the entity information extraction optimization sub-model, the output layer of the event information extraction optimization sub-model, and the output layer of the image association and information source extraction optimization sub-model are associated to obtain the first multi-task collaborative data analysis model; It should be noted that the output layer of the above-mentioned four optimized sub-models, with the feature sharing layer as the core, uses the integrated and constructed feature sharing layer as the underlying general feature base of the entire model. It connects to the independent output layers of the optimization sub-models for time information extraction, entity information extraction, event information extraction, and image association and information source extraction, respectively. The feature sharing layer outputs shared semantic features, which are sent to the output branches of each sub-task. The classification and discrimination logic of each sub-task is retained, forming a network topology structure in which the underlying features are shared and the upper-level task outputs are independent. The multi-task parallel extraction architecture is initially built, resulting in the first multi-task collaborative data analysis model with multi-task synchronous forward inference capability.

[0089] S423, with the goal of minimizing the cross-task loss function, the first multi-task collaborative data analysis model is trained with cross-task collaborative constraints using the training dataset to obtain the second multi-task collaborative data analysis model; The expression for the cross-task loss function is: In the formula, This represents the total loss across tasks, and its value ranges from [0, +]. ); , , , These are the dynamic weight coefficients for the time information extraction subtask, the entity information extraction subtask, the event information extraction subtask, and the image association and information source extraction subtask, respectively. The values ​​are all in the range of (0,1] and the sum of the weights is 1. , , , These are the training errors for the time information extraction subtask, entity information extraction subtask, event information extraction subtask, and image association and information source extraction subtask, respectively, all ranging from [0, +]. ); The preset logical constraint penalty coefficient has a value range of [0, 3, 0.7]. For indicator functions; This is a multi-task logical consistency judgment function; The structured prediction output of the sub-model is optimized for extracting time information. The structured prediction output of the sub-model is optimized for entity information extraction; Optimize the structured prediction output of the sub-model for event information extraction; The structured prediction output of the optimized sub-model for image association and information source extraction.

[0090] It should be noted that the above , , , The deviation between the predicted output of each subtask and the actual labeled label is obtained using the conventional loss calculation method in this field; The preset logical constraint penalty coefficient has a value range of [0, 3, 0.7]; the above indicator function When the condition within the parentheses is true, the value is 1, triggering a penalty mechanism; when the condition within the parentheses is false, the value is 0, and no additional penalty is applied. This is the multi-task logic consistency judgment function. To verify the business logic relevance of the output results of the four core sub-tasks—time information extraction, entity information extraction, event information extraction, image association, and information source extraction—the output is set to 1 if the four outputs are logically consistent and without contradictions; otherwise, a non-1 value is output. The structured prediction output results of the time information extraction optimization sub-model are also included. This includes time category labels and time semantic feature vectors; the structured prediction output of the entity information extraction optimization sub-model. This includes entity category labels and entity semantic feature vectors; the structured prediction output of the event information extraction optimization sub-model. This includes event type, event status label, and event semantic feature vector; and the structured prediction output of the image association and information source extraction optimization sub-model. This includes image identification, location association, and information source feature vectors.

[0091] It should be noted that the aforementioned cross-task collaborative constraint training aims to minimize the cross-task loss function. The training dataset is used to train the first multi-task collaborative data analysis model. Specifically, after inputting the training dataset into the model, a feature sharing layer completes unified feature encoding, outputting structured prediction results for four sub-tasks: time, entity, event, and image association. The loss function is then used to calculate the individual loss for each sub-task. Dynamic weight coefficients are adaptively matched based on the training accuracy of each sub-task. Simultaneously, a logical consistency discriminant function verifies the rationality of the business associations in the time, entity, and event prediction outputs, imposing additional penalty constraints on outputs with logical contradictions. The training process employs a gradient descent algorithm, with minimizing the total loss as the core objective. Forward propagation calculates the loss, and backpropagation iteratively updates the model weights and bias parameters. Multiple rounds of batch iteration and verification are performed until the model's total loss converges and the cross-task logical conflict rate meets the preset requirements. Through cross-task loss weighted optimization and dual constraints of business logic, the model's basic parameters converge and the initial multi-task collaborative capability is fine-tuned, resulting in a second multi-task collaborative data analysis model with the ability to simultaneously extract data from multiple tasks and output logically consistent results.

[0092] It should be noted that the dynamic weight coefficients for the aforementioned time information extraction subtask, entity information extraction subtask, event information extraction subtask, and image association and information source extraction subtask are adaptively and dynamically adjusted during training based on the real-time extraction accuracy and training error of each subtask. Subtasks with large training errors, high recognition difficulty, and low accuracy are assigned greater weights to increase the loss proportion and enhance the model's fit optimization for those subtasks. Subtasks with high accuracy and small errors have their weights appropriately reduced to avoid overfitting. To improve training efficiency, the aforementioned dynamic weight coefficients for the time information extraction subtask, entity information extraction subtask, event information extraction subtask, and image association and information source extraction subtask can also use preset fixed weights, with preferred values ​​of 0.3, 0.3, 0.2, and 0.2, respectively.

[0093] It should be noted that the above-mentioned preset logical constraint penalty coefficient This value controls the penalty intensity when there are logical conflicts in the output results of multiple tasks. The larger the value, the more severe the penalty for logical contradictions, avoiding model outputs that do not conform to business logic. In practical engineering applications, a value of 0.5 is preferred, which can impose appropriate penalty constraints on the business logic conflicts in the prediction results of time, entity, and event, without interfering with the normal training and feature fitting of each sub-task due to excessive penalty intensity, thus ensuring the effect of multi-task collaborative optimization.

[0094] S424, perform iterative collaborative optimization on the second multi-task collaborative data analysis model to obtain the multi-task collaborative data analysis model; It should be noted that the above iterative collaborative optimization process adopts a closed-loop iterative process of "overall training - subtask fine-tuning - overall verification" to carry out iterative collaborative optimization of the second multi-task collaborative data analysis model; the first round of overall training is set to 8 rounds, and the basic learning rate of the model is uniformly adjusted to 5× After each round of overall training, the extraction accuracy of each subtask—time, entity, event, and image association—is evaluated. For subtasks with an accuracy below 90%, independent parameter fine-tuning is performed, with the adjustment range controlled to no more than 30% of the base learning rate. The fine-tuned subtask parameters are then reintegrated into the overall model for further training, and the iterative optimization process is repeated. Iteration stops when the extraction accuracy of each of the four subtasks reaches the preset standard and the overall information extraction F1 value of the model is not less than 92%. Finally, a multi-task collaborative data analysis model with balanced performance of each subtask, close collaboration among multiple tasks, and overall extraction accuracy meets the standard is obtained.

[0095] As can be seen, the unit business record generation method based on a large model described in this embodiment of the invention integrates the feature extraction layers of four sub-models through Transformer, constructs a feature sharing layer, realizes the reuse of underlying features and eliminates parameter redundancy; builds a multi-task parallel architecture, and combines cross-task loss functions and business logic constraints to enhance the synergy of sub-tasks; through iterative optimization of "overall training - sub-task fine-tuning - overall verification", it ensures that the accuracy of each sub-task meets the standard and the overall F1 value is ≥92%, and finally obtains a multi-task collaborative data analysis model with balanced performance, close collaboration and accurate extraction.

[0096] In another optional embodiment, the step of using the unit business classification tag library to train the multi-task collaborative data analysis model to achieve business domain matching capability, thereby obtaining a first unit business chronicle generation model, includes: S431, Perform business domain label normalization processing on the unit business classification label library to obtain a standardized business label set; It should be noted that the above-mentioned business domain tag standardization process involves filtering and deduplicating tags in the unit's business category tag library, unifying tag names, and standardizing tag descriptions to avoid subsequent matching errors due to tag confusion.

[0097] S432, Perform business domain association matching processing on the standardized business label set and the training dataset to obtain a business domain matching training sample set; It should be noted that the above-mentioned business domain association matching process is based on a standardized business tag set. It traverses each text sample in the training dataset and performs association matching between text and business tags from the dimensions of business semantics, topic content, and business affiliation. According to the classification logic of unit business hierarchy, it determines the first-level, second-level, and third-level business domain tags to which each text sample belongs, and completes the association annotation of single text to single tag or multiple tags. At the same time, invalid samples with irrelevant business semantics and low matching correlation are removed, and valid samples with reasonable matching between text and business tags and clear business affiliation are selected and retained. After being sorted and collected, a standardized and usable business domain matching training sample set is formed, which provides qualified sample support for the subsequent training of the model's business domain matching capability.

[0098] S433, using the business domain matching training sample set, the multi-task collaborative data analysis model is trained to perform business semantic adaptation and domain feature fitting to obtain the multi-task analysis model to be adapted. It should be noted that the above-mentioned business semantic adaptation and domain feature fitting training utilizes samples with ambiguous business domain boundaries from the business domain matching training sample set, such as joint project texts involving both "project construction" and "external cooperation." Primary and secondary domain labels and weights are labeled, and the model is trained to learn the domain priority judgment logic. A contrastive learning strategy is adopted, using the above-mentioned business domain matching training sample set to select positive sample pairs of "same type of text - same domain label" and negative sample pairs of "similar text - different domain label" to strengthen the domain discrimination ability. In each round of training, the proportion of ambiguous samples is no less than 25%, accurately learning the correspondence between business labels and text information, fitting the semantic association rules of the unit business domain, completing the deep adaptation of the model to the unit business scenario, and obtaining the multi-task analysis model to be adapted.

[0099] S434, Perform business classification label association constraint and domain matching accuracy calibration on the multi-task analysis model to be adapted to obtain the first unit business record set generation model.

[0100] It should be noted that the above-mentioned business category label association constraint processing strictly follows the unit business category label system to constrain the correspondence between the model output and the labels, avoiding label mismatch and omission issues, and ensuring that the model output accurately matches the business category; the above-mentioned domain matching accuracy calibration processing is to calibrate the deviation in business domain matching, correct problems such as blurred boundaries and unclear domain distinctions, enhance the model's adaptability to business scenarios, and finally obtain the first unit business chronicle generation model that can accurately match the unit's business needs and meet the unit's business chronicle generation requirements.

[0101] As can be seen, the unit business record set generation method based on a large model described in the embodiments of the present invention achieves accurate correspondence between text and business labels by regularizing business labels, accurately matching text and label associations, strengthening training of samples with fuzzy boundaries, and calibrating matching accuracy, thereby improving model adaptability and output accuracy.

[0102] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of a unit business record set generation device based on a large model, as disclosed in an embodiment of the present invention. Figure 2 The described unit business record set generation device based on a large model can be applied to local servers or cloud servers for unit business record set generation, etc., and the embodiments of the present invention are not limited thereto. Figure 2 As shown, this unit business record set generation device based on a large model may include: Unit Information Acquisition Module 201 is used to acquire information on the unit's organizational structure and functional division of labor, as well as the unit's business materials set. The unit business classification module 202 is used to process the unit's organizational structure and functional division information to obtain the unit business classification system and the unit business classification tag library; The unit business material processing module 203 is used to process the unit business material set based on the unit business classification system and the unit business classification label library to obtain a training dataset and a validation dataset. The unit business history set generation model training module 204 is used to train and validate a preset data analysis model using the training dataset and the validation dataset to obtain the unit business history set generation model. The unit business history generation module 205 is used to process the unit business data of the target year using the unit business history generation model to obtain the unit business history of the target year.

[0103] Example 3 Please see Figure 3 , Figure 3 This is a schematic diagram of another unit business record set generation device based on a large model disclosed in an embodiment of the present invention. Wherein, Figure 3 The described unit business record set generation device based on a large model can be applied to local servers or cloud servers for unit business record set generation, etc., and the embodiments of the present invention are not limited thereto. Figure 3 As shown, this unit business record set generation device based on a large model may include: Processor 301; A memory 302 containing executable program code is coupled to the processor 301; The processor 301 calls the executable program code stored in the memory 302 to execute some or all of the steps of the unit business record set generation method based on a large model in Embodiment 1.

[0104] Example 4 This invention discloses a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements some or all of the steps of a unit business record set generation method based on a large model, as described in Embodiment 1.

[0105] The computer-readable storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electronically erasable rewritable read-only memory (EEPROM), read-only optical disc (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0106] Example 5 This invention discloses a computer program product, which includes a computer program stored on a computer-readable storage medium. The computer program includes program instructions, which, when executed by a computer, cause the computer to perform some or all of the steps of a unit business record set generation method based on a large model according to Embodiment 1.

[0107] The system embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0108] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electronically erasable rewritable read-only memory (EEPROM), read-only optical disc (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0109] Finally, it should be noted that the method and apparatus for generating unit business record sets based on a large model disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating unit business record sets based on a large model, characterized in that, The method includes: S1, obtain information on the organization's organizational structure and functional division of labor, as well as the organization's business materials; S2, Process the organizational structure and functional division information of the unit to obtain the unit business classification system and the unit business classification tag library; S3. Based on the unit business classification system and the unit business classification label library, the unit business material set is processed to obtain a training dataset and a validation dataset. S4. Using the training dataset and the validation dataset, the preset data analysis model is trained and validated to obtain the unit business record generation model. The data analysis model includes four independent task-customized neural network models: a time information extraction sub-model, an entity information extraction sub-model, an event information extraction sub-model, and an image association and information source extraction sub-model. Each sub-model includes at least an independent feature extraction layer and an output layer. S5. Using the aforementioned unit business record set generation model, process the unit business data for the target year to obtain the unit business record set for the target year. The step of training and validating a preset data analysis model using the training dataset and the validation dataset to obtain a unit business record generation model includes: S41, using the training dataset, train the time information extraction sub-model, the entity information extraction sub-model, the event information extraction sub-model, and the image association and information source extraction sub-model respectively, to obtain the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model; S42, the Transformer is used to fuse and collaboratively optimize the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model to obtain a multi-task collaborative data analysis model; S43, using the unit business classification tag library, train the business domain matching capability of the multi-task collaborative data analysis model to obtain the first unit business record generation model; S44, perform text data augmentation training on the first unit business record generation model to obtain the second unit business record generation model; S45, perform adversarial optimization training on the second unit business record generation model to obtain the third unit business record generation model; S46, perform cross-format text adaptation training on the third unit business record generation model to obtain the fourth unit business record generation model. S47, Using the verification dataset and preset multi-dimensional verification indicators, a verification experiment is conducted on the fourth unit business record generation model to obtain the compliance verification result; When the compliance verification result is yes, the fourth unit business record set generation model is determined to be a unit business record set generation model; If the compliance verification result is negative, repeat steps S43 to S47 until the compliance verification result is positive.

2. The method for generating unit business record sets based on a large model according to claim 1, characterized in that, The process of processing the organizational structure and functional division information of the unit to obtain the unit's business classification system and the unit's business classification tag library includes: S21, The organizational structure and functional division information of the unit are parsed and processed to obtain the unit's business information; S22, The unit business information is processed using a hierarchical classification method to obtain an initial unit business classification system; S23, The initial unit business classification system is modified to obtain the unit business classification system; S24. Based on the unit business classification system, process the unit business information to obtain a unit business classification tag library.

3. The method for generating unit business record sets based on a large model according to claim 1, characterized in that, The process of processing the unit business material set based on the unit business classification system and the unit business classification label library yields a training dataset and a validation dataset, including: S31, Establish a unit business material index library based on the aforementioned unit business material set; S32, Based on the unit business material index library, perform data extraction processing on the unit business material set to obtain the unit business dataset; S33, Based on the unit business classification system and the unit business classification label library, the unit business dataset is annotated with relevant information of the record set to obtain the unit business labeled dataset; S34, Select and divide the unit business annotation dataset to obtain the training dataset and the validation dataset.

4. The method for generating unit business record sets based on a large model according to claim 1, characterized in that, The step of training the event information extraction sub-model using the training dataset to obtain the corresponding optimized event information extraction sub-model includes: S411, Construct an event type and trigger word mapping library based on the unit's business material set; S412, using the event type and trigger word mapping library, perform event trigger word annotation, event type annotation and event status annotation processing on the training dataset to obtain an event annotation sample set; S413, Perform trigger word context enhancement and sample balancing on the event-labeled sample set to obtain the event training sample set; S414, using the event type training weight calculation model and the event training sample set, the event information extraction sub-model is jointly fitted and trained with event type and event state to obtain the event information extraction sub-model to be optimized; the expression of the event type training weight calculation model is: In the formula, The normalized training weights for the k-th class of event samples take values ​​in the range (0,1). Let be the normalized frequency of the k-th type of event, with a value range of (0,1]. This is a preset frequency enhancement factor, with a value range of [0.3, 2.0]. Let be the sample difficulty coefficient for the k-th type of event. The initial prediction confidence of the event information extraction sub-model for the k-th type of event sample is given, with a value range of [0,1). Let be the rarity coefficient of the k-th type of event. Let k be the number of samples of the k-th type of event. The total number of samples in the training sample set for the event. The preset rarity compensation factor has a value range of [0.3, 2.0]; K is the total number of event types in the event training sample set; S415, perform time and event state correlation verification and feature constraint optimization on the event information extraction sub-model to be optimized to obtain the event information extraction optimization sub-model.

5. The method for generating unit business record sets based on a large model according to claim 1, characterized in that, The process of fusing and collaboratively optimizing the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model using Transformer yields a multi-task collaborative data analysis model, including: S421, the feature extraction layers of the time information extraction optimization sub-model, the entity information extraction optimization sub-model, the event information extraction optimization sub-model, and the image association and information source extraction optimization sub-model are integrated using Transformer to obtain a feature sharing layer; S422, taking the feature sharing layer as the core, the output layer of the time information extraction optimization sub-model, the output layer of the entity information extraction optimization sub-model, the output layer of the event information extraction optimization sub-model, and the output layer of the image association and information source extraction optimization sub-model are associated to obtain the first multi-task collaborative data analysis model; S423, with the goal of minimizing the cross-task loss function, the first multi-task collaborative data analysis model is trained with cross-task collaborative constraints using the training dataset to obtain the second multi-task collaborative data analysis model; The expression for the cross-task loss function is: In the formula, This represents the total loss across tasks, and its value ranges from [0, +]. ); , , , These are the dynamic weight coefficients for the time information extraction subtask, the entity information extraction subtask, the event information extraction subtask, and the image association and information source extraction subtask, respectively. The values ​​are all in the range of (0,1] and the sum of the weights is 1. , , , These are the training errors for the time information extraction subtask, entity information extraction subtask, event information extraction subtask, and image association and information source extraction subtask, respectively, all ranging from [0, +]. ); The preset logical constraint penalty coefficient has a value range of [0, 3, 0.7]. For indicator functions; This is a multi-task logical consistency judgment function; The structured prediction output of the sub-model is optimized for extracting time information. The structured prediction output of the sub-model is optimized for entity information extraction; Optimize the structured prediction output of the sub-model for event information extraction; The structured prediction output of the optimized sub-model for image association and information source extraction; S424, perform iterative collaborative optimization processing on the second multi-task collaborative data analysis model to obtain the multi-task collaborative data analysis model.

6. The method for generating unit business record sets based on a large model according to claim 1, characterized in that, The step of using the unit business classification tag library to train the multi-task collaborative data analysis model to achieve business domain matching capability, thereby obtaining the first unit business chronicle generation model, includes: S431, Perform business domain label normalization processing on the unit business classification label library to obtain a standardized business label set; S432, Perform business domain association matching processing on the standardized business label set and the training dataset to obtain a business domain matching training sample set; S433, using the business domain matching training sample set, the multi-task collaborative data analysis model is trained to perform business semantic adaptation and domain feature fitting to obtain the multi-task analysis model to be adapted. S434, Perform business classification label association constraint and domain matching accuracy calibration on the multi-task analysis model to be adapted to obtain the first unit business record set generation model.

7. A unit business record set generation device based on a large model, applied to the unit business record set generation method based on a large model as described in any one of claims 1-6, characterized in that, The device includes: The unit information acquisition module is used to acquire information on the unit's organizational structure and functional division of responsibilities, as well as the unit's business materials. The unit business classification module is used to process the unit's organizational structure and functional division information to obtain the unit business classification system and the unit business classification tag library; The unit business material processing module is used to process the unit business material set based on the unit business classification system and the unit business classification label library to obtain a training dataset and a validation dataset. The unit business history generation model training module is used to train and validate a preset data analysis model using the training dataset and the validation dataset to obtain the unit business history generation model. The unit business history generation module is used to process the unit business data of the target year using the unit business history generation model to obtain the unit business history of the target year.

8. A unit business record set generation device based on a large model, applied to the unit business record set generation method based on a large model as described in any one of claims 1-6, characterized in that, The device includes: processor; A memory coupled to the processor stores executable program code; The processor calls the executable program code stored in the memory to execute the unit business record generation method based on a large model as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, which, when invoked, are used to execute the unit business record generation method based on a large model as described in any one of claims 1-6.

Images